Note, the lite model in this doc just means MP-CNN without per-dimensional convolution and multiple types of pooling. It uses >8x fewer parameters as the full model!
python main.py sick.mpcnn.lite.model --arch mpcnn_no_per_dim_no_multi_pooling --dataset sick --epochs 19 --lr 0.00086 --regularization 0.0002672 --wide-conv --attention basic --dropout 0.5
| Split |
Pearson's r |
Spearman's p |
| Full model (paper) |
0.8686 |
0.8047 |
| Lite model |
0.8805 |
0.8227 |
python main.py trecqa.mpcnn.lite.model --arch mpcnn_no_per_dim_no_multi_pooling --dataset trecqa --epochs 5 --lr 0.00037 --regularization 0.0017304 --dropout 0
| Split |
MAP |
MRR |
| Full model (paper) |
0.762 |
0.854 |
| Lite model |
0.795 |
0.889 |
python main.py wikiqa.mpcnn.lite.model --arch mpcnn_no_per_dim_no_multi_pooling --dataset wikiqa --epochs 5 --lr 0.00025 --regularization 0.0001088 --wide-conv --attention basic --sparse-features
| Split |
MAP |
MRR |
| Full model (paper) |
0.693 |
0.709 |
| Lite model |
0.696 |
0.708 |